llama2thedog/BIM-JEPA-hf

BIM-JEPA — Colab Demo Bundle

This repository is the all-in-one Colab demo bundle for BIM-JEPA. It is not the canonical model release — it is a self-contained drop used by the Colab demo notebook so that anyone can try BIM-JEPA on a free Colab GPU with zero local setup (no GitHub clone, no PyTorch3D compile).

For the canonical, standalone model releases, please use the dedicated repos:

Repo	Description
llama2thedog/BIM-JEPA-pretrained	Pre-trained foundation encoder
llama2thedog/BIM-JEPA-finetuned-ifcnetcore	Fine-tuned on IFCNetCore (89.37% OA)
llama2thedog/BIM-JEPA-finetuned-bimgeom	Fine-tuned on BIMGEOM (92.43% OA)

How this bundle is used

The demo notebook BIM_JEPA_demo.ipynb in the main repo calls huggingface_hub.snapshot_download on this repo and pulls everything it needs in one shot:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="llama2thedog/BIM-JEPA-hf", local_dir=".")

Contents

File / Folder	Purpose
bimjepa/	The BIM-JEPA Python module (mirror of the GitHub source) so Colab can import bimjepa without cloning the repo
checkpoints/epoch=349-step=60550.ckpt	Fine-tuned IFCNetCore classifier checkpoint (used by the demo)
processed_IFCNetCore_pointclouds_4096_test.zip	Pre-processed IFCNetCore test split (4096-point clouds) for the demo's inference cells
pytorch3d-0.7.9-cp312-cp312-linux_x86_64.whl	Pre-built PyTorch3D wheel for Colab (Python 3.12 / Linux x86_64), so users don't have to compile from source

Paper

Self-supervised learning for BIM element classification using a joint embedding predictive architecture
Jack Wei Lun Shi, Wawan Solihin, Yufeng Weng, Yimin Zhao, Leong Hien Poh, Justin K.W. Yeoh
Automation in Construction

• GitHub repo
• Project page

License

MIT